Abstract: Concerning the limits of the currently used statistic methods of air pollution (not considering correlation and non-orthogonality among forecasting factors results in regression instability and more errors), the linear regression and empirical orthogonal function (EOF) are combined with the stepwise reg ression analysis method, and thus a new fo recasting method in the building forecasting model is proposed.By using this method for forecasting SO₂ density in the heating period, the model fit ting and forecasting show that these models can not only fit the changing tendency of SO₂ density, but also fo recast SO₂ densi ty qui te well, e.g., the grade accuracy is 72.5 percent.In cont rast with the step wise reg ression analysis method, during the forecast experiment, the result of the new forecasting method is more accurate. The new forecasting method has good prospect in application.